On Thu, Mar 21, 2013 at 11:29:37AM -0700, Tejun Heo wrote:
Yes, I actually like that part a lot although I do wish the idle check was inlined.
What I'm wondering is whether the kinda out-of-band decision via sched_select_cpu() is justified given that it can and is likely to go through full scheduling decision anyway. For timer, we don't have that, so it makes sense. For work items, it's a bit different.
To rephrase, *if* the scheduler can't already make proper decisions regarding power consumption on an idlish system, maybe it can be improved to do so? It could as well be that this CPU selection is special enough that it's just better to keep it separate as this patchset proposes. This is something up to the scheduler people. Peter, Ingo, what do you guys think?
Ping. Peter, Ingo?
Viresh, would it be difficult to make another measurement of the same workload with the said workqueues converted to unbound? I think that would at least provide a nice reference point.
Thanks.